Generative & Automated Modelling

Agent-assisted modelling runs along a spectrum of how much the agent decides, and it helps to name three points on it. Parametric modelling is the oldest and most controlled: you build a model governed by rules and parameters - column grid, floor-to-floor, facade module - and changing an input updates the geometry. The rules are yours; the agent, if involved, helps you write, adjust or drive them. Nothing is explored that you did not encode. Script-driven or automated modelling goes a step further: you describe a repetitive or rule-based construction - lay out the parking to this standard, populate this facade, build the stair to these rules - and an agent writes and runs the script that produces it, adapting when a case does not fit. Here the agent automates a build you specified; the recipe is yours, the execution is delegated.

Generative modelling is the most autonomous: you set objectives and constraints - maximise usable area, hit this daylight target, respect these setbacks and this circulation rule - and the agent explores a large space of possible geometries that satisfy the constraints and trade off the objectives, returning many options ranked by how they perform. This is design-space exploration, and it is where agentic modelling is most exciting and most easily misused, because the agent is now proposing forms you did not specify. Across all three, notice that what increases from left to right is the agent's latitude, not your authority: even in full generative mode, you author the frame (goals and constraints) and you make the final choice.

The practical point is to reach for the mode that fits the problem. A great deal of everyday value is in the middle - automating the repetitive, script-shaped modelling that eats studio hours (populating, laying out, standard details) - which is lower-risk than full generative exploration and pays back immediately. Save generative design-space exploration for genuinely open problems - massing, unit mix, layout packing - where surfacing options you would not have drawn is worth the extra discipline of curating them. Match the tool to the task, and remember that more agent latitude always means more curation, not less.

The defining feature of generative modelling is volume: instead of three schemes, you can have three hundred, each scored against your criteria. This is a real gift - it can break a designer out of the first idea that came to mind, reveal that a constraint you thought binding is not, and quantify trade-offs (area versus cost versus daylight versus circulation) that were previously argued by intuition. Used well, a field of options is a map of the solution space that helps you see where the good regions are and why. But it carries a specific danger that every designer using these tools must internalise: a large field of generated options is a field of plausible geometry, not a field of good design, and the volume itself is seductive.

Three failure modes recur. First, plausible-but-wrong: the agent optimises exactly what you told it to and quietly violates what you did not - a layout that maximises area but produces mean, lightless rooms; a massing that hits the daylight target but ignores how the building meets the street. The metric was satisfied; the architecture was not. Second, garbage constraints in, confident options out: if the site model, the setback, the code assumption or the objective you fed it is wrong, every one of the hundreds of options is confidently, uniformly wrong - and the polish of the output hides the error. Third, anchoring on the field: faced with a ranked list, it is tempting to accept the top-scored option because a number blessed it, when the score only measures what you chose to measure and misses everything you did not - beauty, context, buildability, the human experience of the space.

So the skill is to treat the option field as evidence to interrogate, not a verdict to accept. Read why the high-scoring options score well and whether those reasons are the ones that matter. Look hard at what the scoring ignores. Distrust uniformity - if every option shares a flaw, the frame is wrong, not the search. Generation is cheap; discrimination is the expensive, human part, and it is precisely where your fee and your judgement live.

In a generative workflow, the designer's role shifts from making the geometry to two acts that are harder and more valuable: framing the problem and curating the results. Framing is where most of the quality is won or lost, because the agent can only optimise what you express. A good frame states the real objectives (not just the measurable proxies), encodes the true constraints (site, code, budget, program), and is honest about what it leaves out - so you remember to judge those things yourself. A lazy frame - vague objectives, missing constraints, a metric that stands in badly for what you actually want - produces a field that is worthless or, worse, confidently misleading. Time spent sharpening the frame is the highest-leverage time in the whole exercise.

Curation is the second act: reading the returned options with a designer's eye, not just a spreadsheet's. That means understanding the trade-offs the field reveals, identifying the handful of genuinely interesting options (which are often not the top-ranked ones), and asking of each not 'what did it score' but 'is this good, is this right for this place and these people, would I put my name to it'. It also means using the field to learn - to reframe and regenerate as you discover what you actually value - rather than treating the first run as final. The loop is generate, curate, reframe, regenerate; the intelligence is in how you steer it.

And then you decide. The final choice of scheme is a design judgement and a professional act, and it stays with you - the agent proposes, you dispose. An option that a generative run surfaced is still your design when you select, adapt and develop it, and you answer for it exactly as you would for one you drew by hand. This is liberating rather than diminishing: offloading the labour of generating options frees you to spend your judgement where it counts, on framing the problem well and choosing wisely. The generative tool makes you a better-informed decider; it does not make the decision, and it does not carry it.

Generative and automated modelling earns its keep on bounded, well-defined problems where the objectives can be honestly stated and the trade-offs genuinely quantified. Massing studies against area, setback and daylight; unit-mix and layout packing for a residential floor plate; parking and circulation optimisation; feasibility and yield studies early in a project when you want to understand what a site can take. In a small studio, the automated end of the spectrum is often the bigger win - scripting the repetitive modelling that would otherwise burn junior hours - because it pays back immediately and carries less risk than open-ended generation. On the Indian development side, rapid feasibility and yield studies can be genuinely valuable for early client conversations, provided everyone understands they are studies to verify, not designs to build.

The verification discipline is specific to this domain, and skipping it is where designers get burned. Verify the frame first: before you trust any option, confirm the site model, the setbacks, the code assumptions and the objectives are correct - because an error there corrupts the entire field uniformly and invisibly. Check the winners against reality, not just against the metric: does the top-scored massing actually work in section, meet the street, build economically, feel like a place - things the score never measured. Never confuse a feasibility study for a design: a generated option that pencils out on area and cost has not been designed, coordinated, or checked for code compliance and buildability; those remain to be done by you (Modules 4.4, 5.1). And treat any performance number attached to an option (daylight, energy, cost) as a claim to verify with a proper analysis, not a fact, which is exactly the subject of the next lesson.

Held to this discipline, generative modelling is one of the most genuinely creative uses of agents in design - a way to see more of the solution space and choose with better information. Without it, it is a machine for producing confident, buildable-looking wrong answers at scale. The difference, again, is a designer who frames honestly, curates critically, verifies the frame and the winners, and owns the decision.